GatorLM
GatorLM is a custom GPT-style language model developed at the University of Florida.
Model Details
- Architecture: Custom GPT with RoPE, GQA, RMSNorm, and SwiGLU MLP
- Parameters: ~2B
- Context length: 2048 tokens
- Tokenizer: GPT-2 (tiktoken)
- Dtype: bfloat16
Usage
This model uses a custom inference handler. Send requests in the following format:
from huggingface_hub import InferenceClient
client = InferenceClient(model="Krill11/GatorLM1")
response = client.post(json={"turns": [], "message": "Hello, who are you?"})
Input format
{
"turns": [["previous user message", "previous assistant reply"]],
"message": "current user message"
}
turns: list of completed[user, assistant]exchange pairs (empty list for a fresh conversation)message: the new user message
Output format
{
"reply": "GatorLM's response"
}
Training
Fine-tuned via supervised fine-tuning (SFT) on conversational data using the Muon optimizer.